BatchMatMulQueueDescriptor Struct Reference

#include <WorkloadData.hpp>

Inheritance diagram for BatchMatMulQueueDescriptor:

Public Member Functions
void	Validate (const WorkloadInfo &workloadInfo) const
Public Member Functions inherited from QueueDescriptorWithParameters< BatchMatMulDescriptor >
virtual	~QueueDescriptorWithParameters ()=default
Public Member Functions inherited from QueueDescriptor
virtual	~QueueDescriptor ()=default
void	ValidateTensorNumDimensions (const TensorInfo &tensor, std::string const &descName, unsigned int numDimensions, std::string const &tensorName) const
void	ValidateTensorNumDimNumElem (const TensorInfo &tensorInfo, unsigned int numDimension, unsigned int numElements, std::string const &tensorName) const
void	ValidateInputsOutputs (const std::string &descName, unsigned int numExpectedIn, unsigned int numExpectedOut) const
template<typename T >
const T *	GetAdditionalInformation () const

Additional Inherited Members
Public Attributes inherited from QueueDescriptorWithParameters< BatchMatMulDescriptor >
BatchMatMulDescriptor	m_Parameters
Public Attributes inherited from QueueDescriptor
std::vector< ITensorHandle * >	m_Inputs
std::vector< ITensorHandle * >	m_Outputs
void *	m_AdditionalInfoObject
bool	m_AllowExpandedDims = false
Protected Member Functions inherited from QueueDescriptorWithParameters< BatchMatMulDescriptor >
	QueueDescriptorWithParameters ()=default
	QueueDescriptorWithParameters (QueueDescriptorWithParameters const &)=default
QueueDescriptorWithParameters &	operator= (QueueDescriptorWithParameters const &)=default
Protected Member Functions inherited from QueueDescriptor
	QueueDescriptor ()
	QueueDescriptor (QueueDescriptor const &)=default
QueueDescriptor &	operator= (QueueDescriptor const &)=default

Detailed Description

Definition at line 780 of file WorkloadData.hpp.

Member Function Documentation

Validate()

void Validate

(

const WorkloadInfo &

workloadInfo

)

const

Definition at line 4099 of file WorkloadData.cpp.

References armnn::BFloat16, armnn::Float16, armnn::Float32, BatchMatMulDescriptor::GetAxesNotMul(), BatchMatMulDescriptor::GetAxesToMul(), TensorInfo::GetNumDimensions(), BatchMatMulDescriptor::GetPermuteVec(), TensorInfo::GetShape(), WorkloadInfo::m_InputTensorInfos, WorkloadInfo::m_OutputTensorInfos, armnn::NCDHW, armnn::NCHW, armnn::NDHWC, armnn::NHWC, armnnUtils::Permuted(), armnn::QAsymmS8, armnn::QAsymmU8, and armnn::QSymmS16.

{
    const std::string descriptorName{"BatchMatMulDescriptor"};

    ValidateNumInputs(workloadInfo,  descriptorName, 2);
    ValidateNumOutputs(workloadInfo, descriptorName, 1);

    // Inputs must be: both 2D+
    // For inputs X and Y whose dimensions to be multiplied are (M,N) and (I,J) respectively,
    // axes N and I must be the same size

    const auto& inputXInfoBeforeParams = workloadInfo.m_InputTensorInfos[0];
    const auto& inputYInfoBeforeParams = workloadInfo.m_InputTensorInfos[1];
    const auto& outputInfo = workloadInfo.m_OutputTensorInfos[0];
    // Output info has already been inferred

    std::vector<DataType> supportedTypes =
    {
        DataType::BFloat16,
        DataType::Float16,
        DataType::Float32,
        DataType::QAsymmS8,
        DataType::QAsymmU8,
        DataType::QSymmS16
    };

    ValidateDataTypes(inputXInfoBeforeParams, supportedTypes, descriptorName);
    ValidateDataTypes(inputYInfoBeforeParams, supportedTypes, descriptorName);
    ValidateDataTypes(outputInfo, supportedTypes, descriptorName);

    if ((inputXInfoBeforeParams.GetNumDimensions() < 2) ||
        (inputYInfoBeforeParams.GetNumDimensions() < 2))
    {
        throw InvalidArgumentException(descriptorName + ": Input tensors are not 2D or greater.");
    }

    TensorInfo inputXInfoAfterParams;
    TensorInfo inputYInfoAfterParams;

    if((m_Parameters.m_TransposeX && m_Parameters.m_AdjointX) ||
       (m_Parameters.m_TransposeY && m_Parameters.m_AdjointY))
    {
        throw InvalidArgumentException(descriptorName +
            ": Invalid descriptor parameters - Transpose and Adjoint "
            "cannot both be true for a given input tensor.");
    }
    if(m_Parameters.m_TransposeX)
    {
        inputXInfoAfterParams = armnnUtils::Permuted(inputXInfoBeforeParams,
                                                     BatchMatMulDescriptor::GetPermuteVec(
                                                         m_Parameters.m_DataLayoutX,
                                                         inputXInfoBeforeParams.GetShape()));
    }
    else if(m_Parameters.m_AdjointX)
    {
        auto axesToMul = BatchMatMulDescriptor::GetAxesToMul(m_Parameters.m_DataLayoutX,
                                                             inputXInfoBeforeParams.GetShape());
        if(inputXInfoBeforeParams.GetShape()[axesToMul.first] !=
           inputXInfoBeforeParams.GetShape()[axesToMul.second])
        {
            throw InvalidArgumentException(descriptorName +
                ": Adjoint is set to true for input tensor X, but the axes to be adjointed are not square." );
        }
        // Shape remains the same as it's square
        inputXInfoAfterParams = inputXInfoBeforeParams;
    }
    else
    {
        inputXInfoAfterParams = inputXInfoBeforeParams;
    }

    if(m_Parameters.m_TransposeY)
    {
        inputYInfoAfterParams = armnnUtils::Permuted(inputYInfoBeforeParams,
                                                     BatchMatMulDescriptor::GetPermuteVec(
                                                         m_Parameters.m_DataLayoutY,
                                                         inputYInfoBeforeParams.GetShape()));
    }
    else if(m_Parameters.m_AdjointY)
    {
        auto axesToMul = BatchMatMulDescriptor::GetAxesToMul(m_Parameters.m_DataLayoutY,
                                                             inputYInfoBeforeParams.GetShape());
        if(inputYInfoBeforeParams.GetShape()[axesToMul.first] !=
           inputYInfoBeforeParams.GetShape()[axesToMul.second])
        {
            throw InvalidArgumentException(descriptorName +
                ": Adjoint is set to true for input tensor Y, but the axes to be adjointed are not square." );
        }
        // Shape remains the same as it's square
        inputYInfoAfterParams = inputYInfoBeforeParams;
    }
    else
    {
        inputYInfoAfterParams = inputYInfoBeforeParams;
    }

    switch(m_Parameters.m_DataLayoutX)
    {
        case DataLayout::NCDHW:
        case DataLayout::NDHWC:
            if(inputXInfoAfterParams.GetNumDimensions() < 3)
            {
                throw InvalidArgumentException(descriptorName +
                    ": Input tensor X does not have the correct "
                    "number of dimensions for the Data Layout that it has been assigned.");
            }
            break;
        case DataLayout::NCHW:
        case DataLayout::NHWC:
        default:
            break;
    }

    switch(m_Parameters.m_DataLayoutY)
    {
        case DataLayout::NCDHW:
        case DataLayout::NDHWC:
            if(inputYInfoAfterParams.GetNumDimensions() < 3)
            {
                throw InvalidArgumentException(descriptorName +
                    ": Input tensor Y does not have the correct "
                    "number of dimensions for the Data Layout that it has been assigned.");
            }
            break;
        case DataLayout::NCHW:
        case DataLayout::NHWC:
        default:
            break;
    }

    auto axesXToMul = BatchMatMulDescriptor::GetAxesToMul(m_Parameters.m_DataLayoutX,
        inputXInfoAfterParams.GetShape());
    auto axesYToMul = BatchMatMulDescriptor::GetAxesToMul(m_Parameters.m_DataLayoutY,
        inputXInfoBeforeParams.GetShape());

    if(inputXInfoAfterParams.GetShape()[axesXToMul.second]
       != inputYInfoAfterParams.GetShape()[axesYToMul.first])
    {
        throw InvalidArgumentException(descriptorName +
            ": The final axis of input tensor X must be the same size as "
            "the second last axis of input tensor Y.");
    }

    {   // Separate scope so we don't pollute the rest of the scope with our temp variables
        // e.g. NHWC isnt compatible with NCHW as of now
        DataLayout xLayout = m_Parameters.m_DataLayoutX;
        DataLayout yLayout = m_Parameters.m_DataLayoutY;

        if(xLayout == DataLayout::NCHW || xLayout == DataLayout::NCDHW)
        {
            if(yLayout == DataLayout::NHWC || yLayout == DataLayout::NDHWC)
            {
                throw InvalidArgumentException(descriptorName +
                    ": Invalid input tensor data layout combination.");
            }
        }
        if(yLayout == DataLayout::NCHW || yLayout == DataLayout::NCDHW)
        {
            if(xLayout == DataLayout::NHWC || xLayout == DataLayout::NDHWC)
            {
                throw InvalidArgumentException(descriptorName +
                    ": Invalid input tensor data layout combination.");
            }
        }
    }

    // Simulate aligning the ends of the matrix dims and prepending 1's to the beginning of the shorter one
    unsigned int outputTensorDimSize = std::max(inputXInfoAfterParams.GetNumDimensions(),
                                                inputYInfoAfterParams.GetNumDimensions());
    if(outputTensorDimSize-2 > 0)
    {
        TensorInfo tiXNotMul = TensorInfo(TensorShape(outputTensorDimSize-2),
                                          DataType::Float32);
        TensorInfo tiYNotMul = TensorInfo(TensorShape(outputTensorDimSize-2),
                                          DataType::Float32);
        TensorInfo tiOutNotMul = TensorInfo(TensorShape(outputTensorDimSize-2),
                                            DataType::Float32);

        auto doAxisExtension = [&](std::vector<unsigned int> axisIndices, TensorInfo& ti)
        {
            auto sizeDiff = (outputTensorDimSize-2) - axisIndices.size();

            for(unsigned int i = 0; i < sizeDiff; i++)
            {
                axisIndices.insert(axisIndices.begin(), 1);
            }

            for(unsigned int i = 0; i < ti.GetNumDimensions(); i++)
            {
                ti.GetShape()[i] = inputXInfoAfterParams.GetShape()[i];
            }
        };

        auto axesXNotMul = BatchMatMulDescriptor::GetAxesNotMul(m_Parameters.m_DataLayoutX,
                                                                inputXInfoAfterParams.GetShape());
        auto axesYNotMul = BatchMatMulDescriptor::GetAxesNotMul(m_Parameters.m_DataLayoutY,
                                                                inputYInfoAfterParams.GetShape());

        doAxisExtension(axesXNotMul, tiXNotMul);
        doAxisExtension(axesYNotMul, tiYNotMul);

        for(unsigned int i = 0; i < tiOutNotMul.GetNumDimensions(); i++)
        {
            tiOutNotMul.GetShape()[i] = std::max(tiXNotMul.GetShape()[i],
                                                 tiYNotMul.GetShape()[i]);
        }

        ValidateBroadcastTensorShapesMatch(tiXNotMul,
                                           tiYNotMul,
                                           tiOutNotMul,
                                           descriptorName,
                                           "input_X",
                                           "input_Y");
    }
}

---

The documentation for this struct was generated from the following files:

• include/armnn/backends/WorkloadData.hpp
• src/backends/backendsCommon/WorkloadData.cpp